zhaodg / waifu2x-ios

iOS Core ML implementation of waifu2x

waifu2x on iOS

Introduction

This is a Core ML implementation of waifu2x. The target of this project is to run waifu2x models right on iOS devices even without network. For macOS version please refer to waifu2x-mac.

Requirements

• XCode 9+
• iOS 11+

Image format

Images with RGB color space works fine. Others should be converted to RGB before processing otherwise output image will be broken. Alpha channel is scaled using bicubic interpolation. Generally it runs on GPU. It automatically falls back to CPU if image is too large for Metal to process, which is extremely slow. (A bad idea)

About models

This repository includes all the models converted from waifu2x-caffe. If you want to dig into Core ML, it is recommended that you should convert them by yourself.

You can convert pre-trained models to Core ML format and then import them to XCode. The pre-trained model can be obtained from waifu2x-caffe.

You can use the same method described in MobileNet-CoreML. You should not specify any input and output layer in python script.

A working model should have input and output like the following example:

Benchmark

Environment

• iPhone6s - waifu2x-ios on iPhone 6s with iOS 11.1
• iPhone8 - waifu2x-ios on iPhone 8 with iOS 11.0
• iPhone11Pro - waifu2x-ios on iPhone 11 Pro with iOS 13.1 (Neural Engine)
• iPad - waifu2x-ios on iPad Pro 10.5 with iOS 11.1
• PC - waifu2x-caffe on Windows 10 16278 with GTX 960M

Results

All of the tests are running denoise level 2 with scale 2x model on anime-style images from Pixiv.

Test1

Image resolution: 600*849

Device	Time(s)
iPhone6s	6.8
iPhone8	4.0
iPhone11Pro	2.0
iPad	2.9
PC	2.1

Test2

Image resolution: 3000*3328

Device	Time(s)
iPhone6s	129.2
iPhone8	73.5
iPhone11Pro	18.8
iPad	49.2
PC	37.5

Evolution

Device: iPad Image resolution: 3000*3328

Milestone	Time(s)	RAM usage(GB)
Before using upconv models	141.7	1.86
After using upconv models	63.6	1.28
After adding pipeline on output	56.8	1.28
After adding pipeline on prediction	49.2	0.38
Pure MPSCNN implementation*	29.6	1.06

*: With crop size of 384 and double command buffers.

Demo